People cannot distinguish GPT-4 from a human in a Turing test — part 1/3

Cleaned-text code points [0, 15806). Bibliography omitted here; source.md includes it.

# People cannot distinguish GPT-4 from a human in a Turing test

###### Abstract

We evaluated 3 systems (ELIZA, GPT-3.5 and GPT-4) in a randomized, controlled, and preregistered Turing test. Human participants had a 5 minute conversation with either a human or an AI, and judged whether or not they thought their interlocutor was human. GPT-4 was judged to be a human 54% of the time, outperforming ELIZA (22%) but lagging behind actual humans (67%). The results provide the first robust empirical demonstration that any artificial system passes an interactive 2-player Turing test. The results have implications for debates around machine intelligence and, more urgently, suggest that deception by current AI systems may go undetected. Analysis of participants’ strategies and reasoning suggests that stylistic and socio-emotional factors play a larger role in passing the Turing test than traditional notions of intelligence.

## 1 Introduction

### 1.1 The Turing test

Progress in artificial intelligence has led to systems that behave in strikingly humanlike ways. Large Language Models like GPT-4 ( OpenAI, 2023 ) not only produce fluent, naturalistic text, but also perform at parity with humans on a range of language-based tasks ( Chang and Bergen, 2024 ) . These systems are increasingly being deployed to interact with people on the internet, from providing assistance as customer service agents ( Soni, 2023 ) to spreading misinformation on social media ( Zellers et al., 2019 ; Park et al., 2023 ) . As a result, people interacting anonymously online are increasingly forced to ask themselves the question: “Am I speaking to a human or a machine right now?"

Unwittingly, these people are engaging in a real-world analogue of a thought experiment dreamed up three quarters of a century ago by the computer scientist and mathematician Alan Turing. In his seminal article, Turing (1950) proposed a test to measure whether a machine could generate behaviour that was indistinguishable from a human. In his original formulation—which he referred to as the imitation game—a human interrogator would speak to two witnesses (one human and one machine) via a text-only interface. If the interrogator was not able to reliably distinguish between the human and the machine, the machine would be said to have passed ( French, 2000 ) .

Turing’s article “has unquestionably generated more commentary and controversy than any other article in the field of artificial intelligence” ( French, 2000 ) (p. 116). Turing originally envisioned the test as a measure of machine intelligence; if a machine could imitate human behaviour on the gamut of topics available in natural language—from logic to love—on what grounds could we argue that the human is intelligent but the machine is not? However, this idea has accrued a raft of objections in the intervening years, for instance that the test is too easy ( Marcus et al., 2016 ; Gunderson, 1964 ) , or too hard ( Saygin et al., 2000 ) , or too chauvinistic ( French, 2000 ) : a controversy that we return to in the discussion.

Independent of intelligence, the Turing test at its core probes something potentially more urgent—whether people can tell when they are communicating with a machine. Systems that can robustly masquerade as humans could have widespread social and economic consequences ( Frey and Osborne, 2017 ; Zellers et al., 2019 ; Ngo et al., 2023 ) . The Turing test also serves as a window onto our own conceptions of what it is to be human ( Hayes and Ford, 1995 ; Turkle, 2011 ) . As interrogators devise and refine questions, they implicitly reveal their assumptions about what makes humans unique, and which qualities would be hardest to imitate.

Over the last 74 years there have been many attempts to implement Turing tests, though few have been controlled experiments ( Oppy and Dowe, 2021 ) . The Loebner Prize ( Shieber, 1994 ) —an annual competition in which entrant systems tried to fool a panel of expert judges—ran from 1990 to 2020 without deeming a single system to have passed. A recent large-scale study ( Jannai et al., 2023 ) found that humans were 60% accurate in identifying a range of modern language models in two minute online conversations. To date, there have been no controlled experimental demonstrations that any machine has passed the test ( Oppy and Dowe, 2021 ) .

In order to understand whether people are likely to be able to detect deception by current AI systems, we ran a randomised controlled two-player implementation of the Turing test using GPT-4. In our pre-registered hypotheses ( Jones and Bergen, 2024 ) , we predicted that human interrogators would be capable of identifying a baseline system, ELIZA ( Weizenbaum, 1966 ) , but would be at chance in determining whether GPT-4 was a human or a machine.

### 1.2 AI models

We evaluated 3 different types of AI Witness: GPT-4, GPT-3.5, and ELIZA. The first two systems used a hand-crafted prompt to instruct a Large Language Model (GPT-4 or GPT-3.5) how to respond to messages. In a prior, exploratory study ( Jones and Bergen, to appear ) , we tested a wide variety of different prompts—varying the personality, strategy, and linguistic style they instructed the model to adopt. In the present work, we adapted the best performing prompt from the exploratory study for both GPT-4 and GPT-3.5 (for the full prompt, see Appendix A ). The prompt instructed the model to adopt the persona of a young person who was very concise, didn’t take the game very seriously, used slang, and made occasional spelling errors. In addition, the prompt contained general information about the setup of the game, and information about recent news stories that had occurred after the model’s training cutoff. Both models were accessed through the OpenAI API with temperature set to 1. The prompt, alongside the message history in the chat, was used to elicit responses from models. To prevent AI witnesses from responding at superhuman speeds, their messages were sent with a delay based on the number of characters in the response.

A central challenge for interpreting the Turing test is that a system might do well either because it expertly imitates humans, or because the interrogators are very credulous. This concern is often raised under the guise of the ELIZA effect: a phenomenon in which participants attribute human characteristics to even very simple systems ( Marcus et al., 2016 ; Mitchell and Krakauer, 2023 ) . The eponymous ELIZA is a simple rules-based chatbot from the 1960’s that users were found to anthropomorphise ( Weizenbaum, 1966 ) . We included ELIZA as a baseline in our experiment, both to empirically quantify the ELIZA effect, and to serve as a manipulation check in our experiment. To the extent that participants can successfully identify ELIZA but not the GPT-based models as non-human, it suggests that what the latter are doing is nontrivial.

### 1.3 Game setup

We used a two-player formulation of the game, where a single human interrogator conversed with a single witness who was either a human or a machine. While this differs from Turing’s original three-player formulation, it has become a standard operationalisation of the test because it eliminates the confound of the third player’s humanlikeness and is easier to implement ( French, 2000 ; Shieber, 1994 ; Saygin et al., 2000 ) .

Each of 500 participants recruited through Prolific ( prolific.com ) were randomly assigned to one of five groups and played a single round of the game. The first group were human witnesses who were instructed to persuade the interrogator that they were human. The remaining four groups were interrogators who were randomly assigned to question one of the four types of witnesses (GPT-4, GPT-3.5, ELIZA, or Human).

The game interface was designed to look like a conventional messaging app (see Figure 5 ). The interrogator sent the first message and each participant could send only one message at a time. After a time limit of five minutes, the interrogator gave a verdict about whether they thought the witness was a human or an AI, as well as their confidence in and reason for that decision. Finally participants completed a demographic survey that probed individual characteristics hypothesised to affect aptitude at the test. Figure 1 contains examples of games from the study.

## 2 Results

### 2.1 Pass rates

We define the pass rate as the proportion of games in which a given class of witnesses was judged to be human by interrogators. GPT-4 achieved a pass rate of 54%, outperforming GPT-3.5 (50%) and the ELIZA baseline (22%), but lagging behind human participants (67%, see Figure 2).

We conducted analyses to test 5 pre-registered hypotheses ( Jones and Bergen, 2024 ) . First, we hypothesised that humans would not be better than chance at detecting GPT-4. To test this, we fit a logistic regression model to predict pass rate with only an intercept. The intercept was not significantly different from 0 ( z = 0.14 , p = 0.482 z=0.14,p=0.482 ), suggesting that the model’s pass rate was not significantly different from 50%. The 95% credible interval of a Bayesian model predicting the intercept for GPT-4 ranged from -0.24 to 0.55 (equivalent to a pass rate between 44% and 63%). As a manipulation check—to determine whether the design was sufficiently sensitive in principle to detect a difference—we found that the intercept for a model predicting ELIZA’s pass rate was significantly below chance ( z = − 5.24 , p < 0.001 z=-5.24,p<0.001 ), with a credible interval between 15% and 31%.

Results also confirmed our second and third hypotheses: humans’ pass rate was significantly higher than GPT-4’s ( z = 2.42 , p = 0.017 z=2.42,p=0.017 ), and GPT-4’s was higher than ELIZA’s ( z = 4.11 , p < 0.001 z=4.11,p<0.001 ). However, inconsistent with our fourth hypothesis, there was no difference between GPT-4 and GPT-3.5 ( z = 0.497 , p = 0.619 z=0.497,p=0.619 ). Lastly, and again in contrast to our predictions, participants who self-reported as being more knowledgeable about how LLMs work were no more successful in identifying them ( z = − 0.023 , p = 0.869 z=-0.023,p=0.869 ). In exploratory analyses, we found a negative effect of age on interrogator accuracy ( z = − 2.24 , p = 0.025 z=-2.24,p=0.025 ), but no significant effect of any other demographic variables, including gender, level of education, and frequency of interaction with chatbots.

### 2.2 Strategies & reasons

In order to gain qualitative insight into the factors which influenced interrogator decisions, we classified the strategies used by interrogators and the reasons they provided for their verdict. 36% of interrogators engaged in small talk with witnesses, asking about personal details or their daily activities. The second and third most frequent strategy categories were social & emotional (25%)—for instance, asking about the witness’s opinions, experiences, and humour—and knowledge & reasoning (13%), including questions about maths, logic, and general knowledge (see Figure 3).

To estimate the effectiveness of each strategy, we found the mean accuracy of games in which it was deployed. Interrogators were more likely to be accurate when they asked about human experiences (75%), logic & maths (78%), or accused the witness of being an AI (80%), though the latter may result from accusations preceding an accurate verdict rather than the effectiveness of the strategy.

Interrogators’ most frequently provided reasons for their decisions (43%) pertained to linguistic style (e.g. spelling, grammar, capitalization, and tone). 24% related to socio-emotional factors such as sense of humour or personality), and 10% pertained to knowledge and reasoning (e.g. knowing too much or too little).

The most frequent justifications for AI verdicts in Human, GPT-3.5, and GPT-4 games were that the witness appeared to be forcing a persona, was overly informal, or lacked personality. AI verdicts in ELIZA’s games predominantly cited the model’s unnatural responses and lack of contextual awareness. The top reasons for human verdicts were that the witness had a humanlike or informal tone, that their responses were plausible, or that they contained spelling and grammar errors. Classification schemes and more detailed breakdowns of games by reasons and strategies are available in the Appendix.

## 3 Discussion

### 3.1 Does GPT-4 pass the Turing test?

How can we tell when a system has passed the Turing test? In his 1950 paper, Turing predicted that by the end of the century, systems would be able to “play the imitation game so well, that an average interrogator will not have more than a 70 percent chance of making the right identification after five minutes of questioning” (p.442), leading a 30% pass rate to become a target in some discussions ( Oppy and Dowe, 2021 ; Jannai et al., 2023 ) . This benchmark seems arbitrary however, and it’s not clear that Turing meant it as a definition of success ( Saygin et al., 2000 ) .

A baseline of 50% is better justified since it indicates that interrogators are not better than chance at identifying machines ( French, 2000 ) . This definition is especially germane to the goal of discovering whether users can reliably identify other humans in online interactions. One potential issue with this definition of success, however, is that it seems to involve confirming the null hypothesis ( Hayes and Ford, 1995 ) . A system might achieve an accuracy that is statistically indistinguishable from chance because participants were guessing randomly, or because the experiment was underpowered.

In our preregistered analysis, we addressed this problem by using the ELIZA baseline as a manipulation check. Only in the case that our analysis showed a pass rate below chance for ELIZA—indicating that the design is sufficiently sensitive to detect this difference—but not for GPT-4, could the model be judged to have passed. On the basis of this definition, GPT-4 passes this version of the Turing test. Moreover, participants’ confidence scores and decision justifications suggest that they were not randomly guessing: judgments that GPT-4 was human had a mean confidence of 73% (see Figure 2).

At first blush, the low human pass rate could be surprising. If the test measures humanlikeness, should humans not be at 100%? In fact, the human pass rate likely reflects changing assumptions about the quality of AI systems, and is similar to other recent estimates ( Jannai et al., 2023 ) . When AI systems are poor, identifying humans is easy. As interrogators’ confidence in AI systems’ abilities increases, they should become more likely to misidentify humans as AI.

### 3.2 What does the Turing test measure?

Turing originally envisioned the imitation game as a measure of intelligence. A variety of objections have been raised to this idea. Some have objected that the test is too hard ( French, 2000 ) or too chauvunistic ( Saygin et al., 2000 ) , however, these concerns are less pressing if a system does appear to pass ( Turing, 1950 ) . Others have argued that it is too easy. Human interrogators, prone to anthropomorphising, might be fooled by unintelligent systems ( Marcus et al., 2016 ; Gunderson, 1964 ) . We attempted to partially address this concern by including ELIZA as a baseline, but one could always respond that a more stringent or challenging baseline is needed. Still others have argued that no behavioural test can measure intelligence; that intelligence relies upon the right kind of inner mechanism or causal relationship with the world ( Bender and Koller, 2020 ; Block, 1981 ; Searle, 1980 ) (however, see recent philosophical treatments of the potential for LLMs to meet these criteria ( Grindrod, 2024 ; Mollo and Millière, 2023 ; Pavlick, 2023 ) ).